Princeton researchers just released i1, a 3B diffusion model that openly shares its weights, code, and entire data curation pipeline. While not built to outperform massive closed models, it runs easily on consumer hardware and gives anyone a complete, transparent recipe to train image models from scratch.
ZAI just open-sourced SCAIL-2, an end-to-end character animation model that transfers complex motion without relying on ambiguous skeleton maps. It handles intense multi-character combat and can even map human dancing onto flamingos or stylized cartoons. The catch: the full weights clock in at 81GB, so consumer GPUs will likely need to wait for quantized builds.
Rednote just open-sourced dots.tts, a 2B fully continuous end-to-end autoregressive speech system with zero discrete tokens in the pipeline. It achieves top speaker similarity and sub-1% word error rates on Seed-TTS benchmarks, supporting zero-shot cloning and speech editing under an Apache-2.0 license.
Google just released Quantization-Aware Training (QAT) checkpoints for its Gemma 4 models, optimizing them for laptops, phones, and consumer GPUs. By integrating compression directly into the training process, a custom mobile schema squeezes the Gemma 4 E2B model down to under 1GB of memory without sacrificing performance. Why it matters: running high-quality local AI on everyday edge hardware just became drastically lighter on VRAM.
Google just open-sourced Magenta RealTime 2, a live music generation model that runs locally on Apple Silicon Macs with roughly 200ms latency. Controlled via MIDI, text, or audio, it acts like a playable synthesizer rather than a prompt-to-track generator. Why it matters: It shifts generative music from passive offline rendering into an active performance tool for musicians.
Boson AI just released Higgs TTS 3, an open-weights speech model built specifically for real-time conversational agents. Developers can inject inline tags directly into the text stream to tweak emotions, pitch, pacing, and pauses across 100+ languages. It basically turns audio generation into real-time acting direction.
Miso just open-sourced Miso One, an 8B text-to-speech model claimed to be the most emotive voice model in the world. It nails human-like emotional delivery with a lightning-fast 110-millisecond latency—faster than human reaction time. The weights are publicly available now, with an API coming soon for real-time voice apps.
Google just launched Gemma 4 12B, an open-source multimodal model that cuts out separate vision and audio encoders to feed signals directly into the LLM. It nears the performance of their 26B MoE model at less than half the memory footprint, letting developers run local agents on a 16GB laptop.
nex-agi just open-sourced Nex-N2, a new family of agent models post-trained on Qwen 3.5 for real-world coding and long-horizon tasks. Available in 397B and 35B MoE variants activating just 17B and 3B parameters, its adaptive reasoning cuts compute waste by deciding when to think deeply and when to act fast.
TripoAI just open-sourced TripoSplat, a lightweight tool that turns a single 2D photo into up to 260K 3D Gaussians under an MIT license. It runs on minimal code with ComfyUI support, though early tests show it works far better on basic objects than human faces.
JetBrains just released Mellum2, an open-source 12B Mixture-of-Experts model under the Apache 2.0 license. By activating only 2.5B parameters per token, it cuts latency and runs more than 2x faster than similarly sized open models. It is built to handle high-frequency plumbing like routing, RAG compression, and sub-agents across software engineering stacks.
Open-source frontend tool Impeccable just rolled out v3.5 to cure AI-generated design slop, compiling custom guardrails tailored to the specific design flaws of models like GPT-5.5 and Codex. It introduces a beta Live Mode for editing UI directly in the browser via voice or prompt and syncing straight back to source code, powered by a 20x faster CSS detector.
Baidu just open-sourced NAVA, a 6.3B-parameter model that jointly generates 720p video and stereo audio from a single prompt instead of relying on post-production dubbing. It supports reference-audio voice control and reaches state-of-the-art lip-sync benchmarks with up to 5x fewer parameters than rival open baselines.
Researchers from Max Planck and CMU just open-sourced MAMMA, a markerless motion capture system that converts multi-view video into accurate 3D models with only four standard cameras. It tracks complex interactions like dancing and hugging without wearable sensors, easily pulling off Matrix-style bullet time effects. The code and a browser UI are available on GitHub for non-commercial research.
Nvidia just open-sourced SANA-Streaming, a 2B diffusion model that edits minute-long high-res videos in real time while they play. By combining linear recurrent blocks with attention, it hits 24 FPS at 720p on an RTX 5090 while keeping source motion locked in. Code and weights are now up on GitHub and Hugging Face.
Microsoft just released weights for Lens, a 3.8B text-to-image model that rivals larger 6B systems at a fraction of the training cost. Its distilled Turbo variant generates images in just 4 steps and 0.84 seconds on an H100, supporting resolutions up to 1440x1440. It brings high-end, lightning-fast generation directly to independent developers.
NemoStation just open-sourced Marlin-2B, a 2B vision-language model fine-tuned from Qwen 3.5 built to extract structured video data. It answers what happened and exactly when, generating timestamps and boundaries that rival far larger models. Why it matters: you can now run fast, local video search and dense captioning directly on your machine without paying API fees.
Meituan just open-sourced LongCat-Video-Avatar 1.5, a model that turns a single portrait and audio track into realistic talking video. It handles anime and realistic styles, supports multi-person conversations, and cuts diffusion down to 8 steps for a 15x speedup. The 16GB INT8 weights and code are now available on GitHub and Hugging Face.
Researchers just open-sourced SCOPE, an interactive world model for FPS games built on Wan2.2. It supports 10-DoF gamepad and keyboard controls—like moving, aiming, and firing at the same time—and generates consistent 20 FPS video with zero-shot transfer across unseen game styles. Model weights, code, and dataset are live on GitHub.
Perplexity just open-sourced Bumblebee, its internal security scanner built to protect developer machines against supply-chain attacks. Built with Go for macOS and Linux, it checks local packages, editor extensions, and MCP configs entirely in read-only mode—meaning it never runs install scripts or triggers the very malware it hunts for.
Researchers open-sourced ASASR, an image super-resolution tool built on FLUX.1-dev that delivers faithful 4x upscaling while suppressing distorted hallucinated details. It uses adversarial Sobolev DPO alignment and a dual-LoRA setup to preserve true structures. Code, model weights, and a browser demo on Hugging Face Spaces are now available.
Microsoft just rolled out MagenticLite, a new agentic setup designed to run complex workflows across web browsers and local file systems using small models. It pairs a 14B orchestrator called MagenticBrain with Fara1.5, a specialized computer-use model that nearly doubles web navigation performance over its predecessor. It proves you don't need massive frontier models to build capable agents on local hardware.
ByteDance just open-sourced Bernini, a unified video framework pairing an MLLM semantic planner with a DiT renderer. It handles generation and editing across text, images, and video in a single pipeline, putting AI video closer to real directing. Code and model weights are now live on GitHub and Hugging Face.
Tencent just open-sourced Hy-MT2, a multilingual translation model supporting 33 languages in 1.8B, 7B, and 30B MoE sizes. Using AngelSlim 1.25-bit quantization, the 1.8B version shrinks down to just 440MB for smooth on-device use. It prioritizes controllable output like preserving JSON structure and enforcing terminology consistency, making edge-based localization surprisingly practical.
Researchers just dropped Mega ASR, a lightweight speech recognition model built specifically for noisy, distorted real-world audio. Trained on 2.6 million challenging samples covering reverberation, clipping, and severe background interference, it cuts error rates by over 30% against leading baselines. The entire model sits under 5GB, meaning developers can easily run and fine-tune it locally on consumer GPUs.
